结合拆分注意力特征融合的病理图像分割网络  

Pathological Images Segmentation Network Combined Split Attention Feature Fusion

作　　者：邓健志[1,2] 支佩佩 张峰铭 徐国增 田佳 DENG Jian-zhi;ZHI Pei-pei;ZHANG Feng-ming;XU Guo-zeng;TIAN Jia(School of Information Science and Engineering,Guilin University of Technology,Guilin 541004,China;Guangxi Key Laboratory of Embedded Technology and Intelligent System,Guilin University of Technology,Guilin 541004,China;Department of Oncology,Liuzhou People s Hospital,Liuzhou 545006,China;Department of Pathology,Affiliated Hospital of Guilin Medical College,Guilin 541001,China)

机构地区：[1]桂林理工大学信息科学与工程学院,桂林541004 [2]桂林理工大学广西嵌入式技术与智能系统重点实验室,桂林541004 [3]柳州市人民医院肿瘤科,柳州545006 [4]桂林市医学院附属医院病理科,桂林541001

出　　处：《科学技术与工程》2023年第7期2922-2931,共10页Science Technology and Engineering

基　　金：国家自然科学基金(NO81660031);广西自然科学基金(2018GXNSFAA050049)。

摘　　要：针对卷积神经网络在执行病理图像分割任务时,特征提取单一导致分割性能较差的问题,提出了一种结合拆分注意力跨通道特征融合的病理图像分割网络。首先以UNet为基本结构,设计了空洞拆分注意力模块来提取并融合病理图像上细节特征,以增强通道间的特征交互能力,提高分割精度。其次,设计了深度残差幻影模块,在解码特征融合阶段有效获取足够丰富的特征图。最后在公开数据集DSB2018、MoNuSeg上实验,其灵敏度分别为90.13%、89.23%,准确率分别为92.89%、92.51%。为进一步验证算法有效性,将来自合作单位的病理图像自制成数据集ColonCancer,其准确率和灵敏度分别为90.15%、89.94%。实验结果表明,该方法相较于UNet、ResUNet、GhostUNet、TransUNet等算法有效提升了病理图像分割性能,并对实现不同组织病理图像的分割任务具有一定参考价值和意义。In order to solve the problem of poor segmentation performance caused by single feature extraction,when convolutional neural network processing pathological image segmentation task,a pathological image segmentation network based on cross-channel feature fusion of split attention was proposed.Firstly,UNet was used as the basic structure,and a dilated split attention module was designed to extract the detail feature of pathological images,so as to enhance the feature interaction ability between channel and improve the segmentation accuracy.Secondly,the deep residual ghost module was designed to obtain sufficient feature maps effectively in the decoding feature fusion stage.Finally,the experiment was carried out on public datasets DSB2018 and MoNuSeg.The sensitivity are 90.13%、89.23%,and the accuracy are 92.89%、92.51%,respectively.To further verify the effectiveness of the algorithm,the pathological images from the cooperative units were made into ColonCancer dataset.The accuracy and sensitivity of Coloncancer are 90.15%and 89.94%,respectively.The results show that compared with UNet,ResUNet,GhostUNet,TransUNet and other networks,the proposed method can effectively improve the performance of pathological image segmentation,and has certain reference value and significance for the realization of different histopathological image segmentation tasks.

关 键 词：图像分割 拆分注意力 深度可分离 病理图像 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]